Unintended Consequences of Free College:Self-Selection into the Teaching Profession∗

Rosa Castro-ZarzurUniversity of Maryland

Ricardo EspinozaOECD

Miguel SarzosaPurdue University

November 5, 2019

Abstract

Teacher quality is one of the most relevant factors influencing student learning.However, attracting and retaining skilled people to the teaching profession is challenging.In this paper, we study how making college tuition free affects the pool of studentspursuing a teaching career. We exploit the conjunction of two tuition-financing policiesimplemented in Chile: a scholarship introduced in 2011 for teaching majors, and amassive 2016 reform that made college tuition free for students from households in thebottom 50% of the income distribution. We use the programs’ differences in timing andeligibility criteria to study the effects free college had on the self-selection of students intoteaching majors. We find that free college decreased the relative returns to pursuinga teaching career, making it substantially less popular among relatively poor high-performing students who now self-select into degrees with higher returns. We find thatthe reform reduced the academic qualifications of the pool of students entering theteaching programs, which can negatively affect long-term teacher quality.

JEL Classification: I22, I24, I26Keywords: Teaching profession, free college, career choice, self-selection, access to tertiaryeducation∗We would like to thank the following for comments and suggestions: Joe Altonji, Felipe Barrera, Hugo

Ñopo, Jean Lafortune, Snaebjorn Gunnsteinsson, Fabio Sánchez, Fernando Saltiel, Guido Matías Cortés,Alejandra Ramos, Catherine Rodríguez, Andrés Moya, Diana Hincapié, Suzanne Duryea, Sebastián EspinozaRojas, Justin Tobias, Kevin Mumford, and seminar participants at DNP (2018, Bogotá, Colombia), Pur-due University (2018, West Lafayette, IN) and Inter-American Development Bank (2018, Washington, DC),LACEA (2018, Guayaquil, Ecuador), SoLE (2019, Arlington, VA), Pontificia Universidad Católica (2019,Santiago, Chile), OECD Applied Economics Seminar (2019, Paris, France), 2nd IZA/World Bank/NJD Con-ference on Jobs and Development (2019, Washington, DC), ESPE (2019, Bath, UK), IWAEE (2019, Catan-zaro, Italy), APPAM International (2019, Barcelona, Spain) and the 4th IZA Workshop on the Economics ofEducation (2019, Bonn, Germany). The views expressed in this paper are those of the authors and do notnecessarily reflect those of the OECD or of its member countries. Castro-Zarzur: rcz@umd.edu; Espinoza:ricardo.espinoza@oecd.org; Sarzosa: msarzosa@purdue.edu.

1

1 Introduction

Teachers are a critical input in the formation of early human capital. Teacher quality is one

of the most relevant external factors influencing academic achievement.1 Effective teachers

can create fruitful learning environments, inspire and motivate students, compensate for the

lack of a favorable home environment, and help level the playing field for students who lag

behind. A seminal study in the US shows that having a higher-quality teacher raises the

probability of college attendance, the quality of the colleges to which students attend, and

annual earnings (Chetty et al., 2014).

But, where do good teachers come from? There is convincing evidence that most high-

quality teachers were once top- performing high school graduates.2 Therefore, attracting high

academically performing students to teaching majors is critical for the success of education

systems. Thus, policies intended to attract, retain, and develop academically talented people

for the teaching profession have become central to education policy (OECD, 2005; Bruns and

Luque, 2015).3 However, in most countries, colleges struggle to attract students to teaching

majors, particularly high performers. For example in the US, teacher education enrollments

dropped from 691,000 to 451,000 between 2009 and 2014, a 35% reduction (Sutcher et al.,

2016); and for both, Latin America and the United States, there is robust evidence showing1Several studies have found positive and meaningful impacts of teacher performance on student learning

outcomes and academic achievement in kindergarten (Araujo et al., 2016), elementary school (Schacter andThum, 2004; Bau and Das, 2019), and high school (Hanushek et al., 2018). Moreover, some evidence alsosuggests that teacher effects are persistent and matter not only in the current or following year but also insubsequent years (Konstantopoulos, 2011).

2Evidence from Teach for America (TFA), a program that recruits graduates from selective colleges inthe US to teach in the most challenging K-12 schools, shows that students lectured by TFA teachers scorehigher on standardized tests despite the lack of experience of TFA teachers (Glazerman et al., 2005; Xuet al., 2011). Similarly, impact estimates of Enseña Chile, the Chilean adaptation of TFA, suggest thatplacing outstanding college graduates in the most vulnerable schools results in significant student gains incognitive and non-cognitive abilities (Alfonso et al., 2010). Furthermore, highly successful education systemslike Finland, Singapore, and South Korea rely on extremely competitive processes to select those allowed tobecome teachers before they enroll in teaching degrees (Auguste et al., 2010; Seng Tang, 2015).

3Such policies typically combine monetary and non-monetary incentives in order to increase the returnof pursuing a teaching degree and to make it more appealing to students who may choose more rewardingdegrees in other fields (OECD, 2005). These efforts include special scholarships, offering free study-abroadsemesters, stipends, and allowances to cover living expenses (Santiago, 2002; Claro et al., 2013).

2

that students enrolling in teaching degrees are disproportionately drawn from the lower end of

the academic proficiency distribution (Santiago, 2002; Eide et al., 2004; Balcázar and Ñopo,

2014), and tend to score lower on IQ tests than those who enroll in other fields (Lang and

Palacios, 2018). This can result in a misallocation of talent if one considers the production

of human capital of future generations.

Economic literature agrees that, to a large extent, this state of affairs is explained by a

combination of low expected labor market returns and low social recognition of the teaching

profession (Elacqua et al., 2018; OECD, 2018). Teachers are among the lowest paid college

graduates.4 Recent evidence shows that the economic returns to teaching degrees is relatively

low or even negative, meaning that teacher earnings may not compensate the financial and

opportunity costs of pursuing such degrees, and that high academically performing students

systematically enroll in degrees with higher economic returns, such as STEM, business, and

law (Gonzalez-Velosa et al., 2015; Espinoza and Urzua, 2016; Hastings et al., 2013).

Despite the consensus on the causes of the negative correlation between academic achievement

and the likelihood of enrolling in teaching majors, economic literature has yet to explore how

the problem could potentially be exacerbated by the increasing pressure citizens have put

on governments and education systems to reduce or eliminate financial constraints affecting

access to tertiary education. Policies aimed at making access to tertiary education more

affordable and equitable will not only alleviate the financial burden of pursuing a college

degree, but will also affect how students self-sort into different majors (Bucarey, 2018). In

this paper, we study how tuition-free college affects the pool of students pursuing a teaching

career. We exploit a major 2016 reform carried out in Chile that made college tuition free

for students from households in the lower 50% of the income distribution. We leverage the

fact that the introduction of free college affected the application behavior of distinct groups

of students depending on their eligibility to the Beca Vocacion de Profesor (BVP) tuition4In OECD countries, for example, primary teachers are paid 85% compared to other tertiary-educated

workers (OECD, 2017a). Studies in middle-income countries find that this effect is still significant aftercontrolling for observable characteristics typically linked to labor productivity (Mizala and Ñopo, 2016).

3

grant, a scholarship program implemented in 2011 that was successful in bringing high-quality

students into teaching majors (Castro-Zarzur, 2018). The BVP granted full scholarships to

students willing to enroll in teaching majors who scored in the top 30% of the college entrance

exam. The free-college policy equalizes the relative prices of studying a wide range of different

majors, potentially offsetting the incentives set out by the BVP.

Using a difference-in-difference strategy on a rich administrative dataset containing test

scores, student applications, and enrollment, we compare the application and enrollment

behavior of students in cohorts before and after the implementation of the reform. Thus,

we identify the causal effect of eliminating tuition fees on student preferences for teaching

degrees and the extent to which it affected the academic qualifications of students pursuing

a teaching career. Our results suggest that granting tuition-free access to college decreased

the demand for teaching programs of top-performing students. In consequence, students ad-

mitted to teaching degrees achieve, on average, lower academic performance than before the

policy. The introduction of tuition-free access to college reduced the probability of applying

to a teaching major among high-performing students who come from relatively poor back-

grounds by about 15.8%, offsetting the gains obtained by the BVP scholarship during its

first years of implementation. Furthermore, the average academic proficiency among those

who applied to a teaching degree fell by 14% of a standard deviation in math and 8.8% of

a standard deviation in language, while the average score of those who were accepted into

other majors remained unchanged. The drop was concentrated on the relatively poor high

school graduates whose score fell by around 16% of a standard deviation in both math and

language.

Our paper highlights the potential unintended consequences of policies distorting equilibrium

prices in markets such as higher education, and the potential long-term effects of such policies

on teacher quality through a decrease in the qualifications of students pursuing teaching

degree programs. Thus, we provide an important input to the ongoing international debate

4

on free college as we present evidence of some negative—albeit unintended—consequences

of such policy, going beyond the commonly used argument against the policy: its financial

burden on governments.

Our findings are likely to be informative even in contexts lacking previous ability-based

scholarships, such as the BVP. Teaching majors tend to be the least costly in many countries—

including the United States, where differential tuitions are being increasingly used. Hence,

in the absence of tuition-free college, the relatively cheap teaching majors are a pathway to

higher education for low-income students. As free college equalizes the tuition cost across all

majors, this allows high performers from poorer backgrounds to substitute relatively cheap

majors with other—likely costlier—programs. In that sense, the BVP is just a vehicle for

identification as it allows us to implement a difference-in-difference strategy while providing

important insights on the potential impact of free college on self-selection into the teaching

profession in more general settings like the US, where the college tuition-free policy is gaining

momentum and several existing programs are already pushing in that direction (e.g., New

Mexico state-wide free college policy and the Excelsior program in New York).

The paper is structured as follows. Section 2 gives an overview of the institutional background

of the teaching profession in Chile and explains the recent reforms of the higher education

system. Sections 3 and 4 present the data and empirical strategy. Section 5 presents the

main results and section 7 discusses the results and concludes.

2 Institutional Background

2.1 Access to Tertiary Education in Chile

Access to higher education has expanded considerably in Chile during the last 25 years (Es-

pinoza and Urzua, 2016). The number of students enrolled in undergraduate degree programs

5

has increased fourfold, from less than 250,000 in 1990 to 1.2 million in 2016 (Centro de Es-

tudios MINEDUC, 2017). According to the World Bank Data, Chilean net enrollment rate

is 90.3%, making it the fourth-highest rate in the world. Such a high degree of coverage is

the product of a multi-tiered system that comprises three types of tertiary education insti-

tutions: Universities (Universidades) serving 42% of the enrolled population, Professional

Institutes (PI, Institutos Profesionales) with 31% of the enrollees, and Technical Training

Centers (TTC, Centros de Formación Técnica) enrolling the remaining 11% (Centro de Es-

tudios MINEDUC, 2017). TTCs offer mostly two-year vocational programs and universities

and PIs more often offer four to five-year majors. As opposed to PIs, universities are research-

oriented institutions and have the exclusive right to offer certain degree programs such as

medicine, law, and teaching degrees. Universities are further divided into two categories: 1)

the 27 “traditional universities," which are part of the Consejo de Rectores de Universidades

Chilenas (CRUCH)—a consortium that encompasses public and private universities founded

before 1981; and 2) the remaining “non-traditional universities," all private and founded after

1981.

Student admissions to universities are determined based on merit within a dual system.

First, a group of 41 universities, comprising all “traditional universities” and 14 private in-

stitutions, run a centralized matching system in which students apply by listing and ranking

their preferred career choices, understood as a unique program-university pairs (e.g., teaching

in university X, medicine in university Y, etc.). Each student can list up to 10 pairs. After

the application process, students are matched with their choices using a deferred-acceptance

matching algorithm determined upon the scores obtained in a nationwide standardized uni-

versity entrance exam called Prueba de Selección Universitaria (PSU) and high school grades

(Espinoza et al., 2017). The remaining universities run a decentralized process, although PSU

scores are still required for admission. Because of the lack of a coordinated assignment in

these remaining universities, students are not selected purely on merit. Some of these uni-

versities accept students on a first-come, first-served basis. Finally, PIs and TTCs also run

6

Table 1: Characteristics of the Teaching Degrees by Type of Institution

Universities ProfessionalCentralized Decentralized Institutes (PI)

# of first-year students† 7,606 13,633 1,490# of programs 295 702 112Annual tuition (2015 USD) 2,971.0 2,427.9 1,712.5

Note: Unweighted average tuition across programs is shown. †Number of students calculatedbased on the number of non-zero vacancies opened by each program reported to SIES. PIsonly offer assistant teacher diplomas. Source: SIES/CNED, 2015.

non-selective and decentralized admission processes.

2.2 The Teaching Profession in Chile

Chilean legislation requires that pre-primary and primary teachers hold a teaching degree

granted by a state-recognized university. PIs and TTCs are allowed to offer only assistant

teacher diplomas. There are two main ways students can enter the teaching profession. They

can enroll in a teaching degree holding a high school diploma, in which case they obtain

their degree in four to five years. Otherwise, students already holding a bachelor (or higher)

degree can enroll in programs that train and certify professionals to teach in schools. These

programs are shorter (one to two years) and give students pedagogical training.

In this paper, we focus on the incentives that delimit the pathway between high school

graduation and teaching degrees. Hence, our analysis zeroes in on the undergraduate teaching

majors and excludes postgraduate teaching degrees. In this context, Table 1 shows that

in 2015 there were 1,109 four-to five-year undergraduate teaching programs, out of which

around a fourth are provided by CRUCH universities.5 These programs are the ones that

enjoy greater prestige as evidenced by the fact that they charge on average substantially

higher tuition than teaching programs in non-CRUCH universities.5We define an undergraduate teaching program as one that grants a BA, that is deemed as education-

related by the UNESCO classification, and requires at least 8 semesters for completion.

7

Table 2: Characteristics of Undergraduate Degrees Offered by Type of Degree

Type of degreeTeaching STEM Others

Number of programs 256 384 663Duration (years) 4.65 (0.03) 5.21 (0.04) 4.97 (0.02)Annual tuition (2015 USD) 3,273.81 (43.83) 4,602.01 (48.85) 4,843.54 (56.16)

PSU score among enrolleesMean 567.61 (2.08) 588.44 (2.42) 581.65 (2.08)Min 505.20 (1.80) 503.69 (2.31) 502.32 (2.15)Max 660.31 (3.71) 697.79 (3.03) 685.15 (2.26)

Note: The unit of observation is a degree-university pair. The sample includes all degrees offered inuniversities in centralized matching system. We exclude from the sample degrees offered by ProfessionalInstitutes and universities not in the centralized admission process. Standard deviation in parenthesis.Source: DEMRE and SIES, 2015.

Table 2, on the other hand, compares the four-to-five-year undergraduate teaching programs

with programs in STEM and non-STEM fields granted by universities in the centralized

admissions system. We see that teaching degree programs tend to be shorter than STEM

degree programs, by about a semester, and their yearly tuition is on average 29% cheaper

than STEM majors and 32% cheaper than non-STEM majors. Table 2 also shows that

at the university-major level non-education majors are more selective in terms of PSU score

requirements. Interestingly, the gap observed between the average teaching and non-teaching

university majors comes from the high-achievers and not from the low-scorers. While there

is virtually no difference between the typical minimum scores across types of majors, the

typical maximum score in a STEM major is 37.48 points greater than in a teaching major.

That is around a third of a standard deviation, as the PSU is designed to have a mean of

500 and a standard deviation of 110.

As in many other parts of the world including the US, Chile’s teaching programs struggle

to attract the most talented students (Eide et al., 2004; OECD, 2005; Alvarado et al., 2012;

Balcázar and Ñopo, 2014; Castro-Zarzur, 2018; Gomez et al., 2019). Figure 1 shows a strong

negative relation between early math proficiency—as measured by a nationwide assessment in

8

Figure 1: Fraction of Students Applying a Teaching Major by 4th Grade Math Score, 2015Graduating Cohort

(a) Listing a Teaching Major in the Choice Set

.1.1

5.2

.25

% A

pp

lyin

g t

o a

Te

ach

ing

Ma

jor

in 2

01

5

0 20 40 60 80 1004th Grade Math Score Decile

(b) Listing a Teaching Major as Top Choice

.04

.06

.08

.1.1

2.1

4%

Ap

ply

ing

to

a T

ea

ch

ing

Ma

jor

in 2

01

5

0 20 40 60 80 1004th Grade Math Score Decile

Note: The left figure shows the share of students from the 2015 graduating cohort who listed a teachingmajor anywhere in their choice set by the 4th grade math SIMCE decile. The right figure shows the share ofstudents from the 2015 graduating cohort who listed a teaching major as their top choice by the 4th grademath SIMCE decile. The lines in both figures represent a fitted regression line.

Table 3: Characteristics of Undergraduate Freshmen by Type of Degree

Type of degreeTeaching STEM Others

Overall PSU score 574.88 (0.571) 608.37 (0.474) 596.43 (0.332)Language PSU score 587.43 (0.722) 586.59 (0.527) 601.65 (0.387)Math PSU score 562.34 (0.700) 630.14 (0.530) 591.20 (0.373)% From top 10% PSU score 12.32 (0.362) 32.83 (0.315) 26.89 (0.215)% Female 62.98 (0.531) 27.47 (0.299) 59.10 (0.239)Mother’s years of schooling 5.99 (0.028) 6.56 (0.018) 6.74 (0.013)% from public schools 29.32 (0.501) 24.43 (0.288) 22.79 (0.204)% from voucher schools 58.87 (0.541) 55.10 (0.334) 50.12 (0.243)% from private schools 11.26 (0.348) 20.02 (0.268) 26.44 (0.214)% Teaching as 1st choice 20.16 (0.214) 0.09 (0.009) 0.21 (0.010)% Teaching in top 3 choices 47.90 (0.267) 0.67 (0.025) 1.63 (0.028)Number of students 8,233 22,187 42,358

Note: The sample includes all students accepted in university degrees under the centralized matching system. Programsconsidered require at least 8 semesters for completion. We exclude from the sample students enrolling in ProfessionalInstitutes and universities running decentralized admission processes. Standard error in parenthesis. STEM programsinclude degrees in the fields of Sciences and Engineering according to the UNESCO classification of degrees. Source:DEMRE and SIES, 2015.

9

4th grade in elementary school, when they were 10 years old—and the probability of applying

to a teaching major. Figure 1a indicates that around 22% of low-scoring children as measured

at age 10 will list a teaching major when they apply to college eight years later. That figure

drops to 11% for top-scoring students. In the same way, Figure 1b reveals that while 11%

of low-scoring 10-year-olds will list a teaching major as their top choice when they reach

college age, only 5% of top scoring students will do so. Table 3 compares the characteristics

of students enrolled in teaching degrees with those enrolled in STEM and other disciplines.

We find that the gap in the average PSU score between those enrolled in teaching and STEM

majors is about 30% of a standard deviation in favor of the latter. The students who end

up enrolled in teaching degrees even trail other non-STEM majors (e.g., liberal arts, social

sciences) by more than a tenth of a standard deviation. The gaps come mainly from the

difference in math proficiency. STEM major enrollees score on average 61.6% of a standard

deviation more than teaching major enrollees on the math section of the test.6 Furthermore,

of those who pursue a teaching degree, only 12.3% come from the top 10% of the PSU

distribution, while a third of those enrolled in STEM do. The gaps reflect substantial low

desirability and competitiveness of enrolling in teaching majors relative to others. This is

made evident by the fact that only a fifth of those who enrolled in teaching programs had

teaching as their top career choice, while virtually none of those who ended up in majors

other than teaching listed teaching as one of their top three choices. Table 3 also shows

that teaching is overwhelmingly female dominated. Two-thirds of the entrants to teaching

majors were females, while only 25.7% of those going into STEM majors were. Teaching

major enrollees are relatively more disadvantaged high school graduates: more of them come

from public schools, and their average mother’s education is half a year and three-fourths of

a year less than STEM and non-STEM majors enrollees, respectively.6This relates to the fact that, when tested in the TEDS-M—an international study that quantifies the

math proficiency of those who are studying to become math teachers—future Chilean high school mathteachers ranked second to last, while those who would become elementary school teachers ranked last, behindcountries that are vastly less developed, like Botswana and Philippines (Elacqua et al., 2018).

10

The reasons behind the low demand for teaching degrees among relatively skilled students

and those coming from affluent backgrounds are multiple and intertwined. First, there is

low social recognition of teaching as a profession (Elacqua et al., 2018). According to recent

surveys, only a third of parents would like their children to be teachers, and two-thirds of

teenagers consider the teaching profession to be among the least prestigious occupations,

with dismal favorability numbers similar to those of music and theater (Cabezas and Claro,

2011). Second, teachers are underpaid (Mizala and Ñopo, 2016) and labor market returns are

very low, even though tuition is also relatively low (see Table 2). Chilean teachers earn on

average 56% of what comparable college graduates do (Elacqua et al., 2018). Thus, teachers’

lifetime incomes are not very different from people without a college degree (Espinoza and

Urzua, 2016). In fact, Gonzalez-Velosa et al. (2015) estimate that Chilean education majors

have a lifetime rate of return of -1%. Third, career development and salary raises are based

on seniority and not on merit. Despite these downsides, the teaching profession is still

attractive to many individuals, particularly those who are relatively more risk averse (Lang

and Palacios, 2018) and therefore value income and job stability, and also to those who prefer

longer holiday periods.

As a response to this worrying picture, the Chilean government has deployed a number of

attempts to reform the teaching profession and to make it more competitive and appealing

to prospective students (OECD, 2017b; Santiago et al., 2013). These initiatives include

offering higher remuneration to teachers, reforming and expanding professional development

opportunities, and subsidizing or fully paying the tuition of teaching degrees. Probably the

most well-know initiative, the “Teaching Calling Scholarship” (Beca Vocacion de Profesor)

or BVP, was introduced in 2011. The BVP recruits students scoring within the top 30% of

the PSU into accredited, high-quality teaching programs. It is a merit-based scholarship and

does not select student based on their income. It is available to prospective college freshmen

scoring 600 or more in the PSU exam enrolling in a teaching major that participates in the

11

program.7 The scholarship has different benefits depending on the student’s score. For those

scoring on or above 600, it covers full tuition fees for the duration of the degree. Those who

score on or above 700 are also eligible for a monthly stipend of about USD$150. Finally,

for those scoring 720, the BVP offers a one-semester exchange program in a university in a

foreign country (Bonomelli, 2017).

Until the introduction of the free-college program, the BVP was the only public scholarship

covering 100% of tuition fees and had shown to be an effective mechanism to increase the

participation of top scoring students in teaching degrees. Alvarado et al. (2012) find that in

the year of its introduction, the BVP increased the probability of applying to and enrolling

in teaching programs for eligible students (i.e., relatively better students in terms of PSU),

and the share of students coming from the top-third of the PSU distribution increased from

10.7% to 18.1%. Claro et al. (2013) analyze the effect of the BVP on the profile of the

students admitted to education programs in 2011 at CRUCH universities. They find that

with the introduction of the scholarship, the probability of applying and being admitted

to teaching programs significantly increased for better-performing students who come from

poorer households and have relatively undereducated mothers. Finally, Castro-Zarzur (2018)

studies the impact of the BVP during 2011-2016, a period in which several older tuition

scholarship programs were expanded and the free college reform was introduced. She finds

that the scholarship had a positive local impact on both applications to and enrollment in

teaching majors in the first years following its introduction; however, after 2013 the impact

tends to diminish, and it is no longer significant by 2016 (year of the introduction of the free

college reform). Consistent with these findings, Castro-Zarzur also reports that both eligible

applications and accepted BVP scholarships decreased by 56% between 2011 and 2016, which

suggests an important decline in the interest in this grant by students from newer cohorts,

particularly those coming from public schools.7BVP programs are teaching majors that are quality-accredited, require full-time dedication, and only

admit students with PSU ≥ 500.

12

2.3 The Tuition-Free College Policy

Starting in 2011, a series of countrywide student-led protests demanded more financial sup-

port from public sources.8 Protestors demanded the repeal of the current market-based

system in which universities compete with each other for funding and students. Based on

the idea that education is a fundamental human right, their aim was to replace the prevail-

ing structure with a publicly run system where the state provides free and universal higher

education.9 However, it was not until 2014 that the Chilean government pushed for major

reforms, including one aimed at making college tuition-free. The main goal of the reforms

was to make access more equitable so that all students, independent of their family income,

had access to a high-quality and publicly funded higher education.

The complexity of the system, characterized by the coexistence of private and public univer-

sities (all charging tuition fees) and the high cost of the reform, made it difficult to implement

(Espinoza and Urzua, 2015). In 2016, the government launched a partial version of a universal

tuition-free program. From the universities’ perspective, participation was non-mandatory.

All 16 public universities automatically joined the program and 14 private institutions de-

cided to participate.10 Tuition was abolished for students from families from the bottom

50% of the income distribution enrolling in participating universities. Universities receive a8The Chilean government funded the higher education system through three main mechanisms. First, it

provides unconditional lump-sum transfers to CRUCH universities. Second, it finances research activities ona competitive basis. Most of these funds are available to all quality-accredited institutions. Finally, it fundsstudents through scholarships and state-subsidized loans (see Beyer et al., 2015). These demand-side subsidieshave increased dramatically over the past decade, especially after the introduction of a large state-guaranteedloan system in 2006 (see Figure AF1 in the Appendix).

9The government’s first response was to expand the two largest college scholarship programs that existedby that time: the Beca Bicentenario (BBIC) available for those going to college at CRUCH universities,and the Beca Juan Gomez Millas (BJGM) available to students attending either CRUCH universities orany quality-accredited higher education institution. Between 2011 and 2014, these two tuition grants—whichcover around 80% of the total tuition costs—gradually went from being available for students with PSU ≥ 550who came from the first two income quintiles, to include students with PSU ≥ 500 coming from deciles oneto seven of the income distribution. Tables A4 in the Appendix show that the expansion of these scholarshipswas finalized before the time period we explore in this paper.

10Professional Institutes (PIs) and Technical Training Centers (TTCs) were excluded from the tuition-freeprogram.

13

regulated per-student fee.11 And knowing that universities would have strong incentives to

drastically change the number of places they made available in each program as a response

to the elimination of tuition (Dynarski, 2003; Abraham and Clark, 2006), the government,

by law, forbid it. Table A5 in the Appendix shows that universities complied, as we see no

change in the number of programs offered, the number of places made available, or in the

number of first year enrollees.

3 Data

We use three different administrative datasets in this study. First, we use data on student

applications to universities in the centralized admission system. These data are managed and

maintained by the Departamento de Evaluación, Medición y Registro Educacional (DEMRE),

an entity that is part of Universidad de Chile, the main public university and one of the 33

higher education institutions affiliated to the centralized admission system as of 2016. The

DEMRE designs the PSU test and administers it nationwide. The PSU evaluates students

in four subjects: math, language, science, and social sciences. Scores range from 150 to 850

with a mean of 500 and standard deviation of 110. The PSU dataset contains individual

information on PSU score, as well as demographics such as gender and age, and self-reported

socioeconomic characteristics such as family income (discrete categories), household size, and

parental education. It also contains students’ high school GPA, their year of graduation and

the school they graduated from and its nature (i.e., public, voucher, private). Students take

the PSU at the end of high school but it is relatively common for individuals to retake the

exam. We restrict our sample to students who graduated from high school the year before

entering higher education.

For those who apply to the centralized admissions system after taking the PSU exam, we11The per-student transfer is defined by a technical body, and is calculated using a complex formula that

takes into account quality and efficiency indicators. See Gobierno de Chile (2017) for more details.

14

also observe the ranking list of up to 10 major-university pairs submitted by each student in

each year. Students and programs are matched using a deferred-acceptance algorithm, which

takes into account student rankings, programs’ preferences (students with higher scores are

preferred to student with lower scores), and quotas. Therefore, students may not be matched

with their most preferred option. The dataset reports the outcome of the application process.

Specifically, we observe the acceptance/rejection decision to each program students applied

to.

Second, we use information on the supply of higher education programs, provided by the

National Education Council (Consejo Nacional de Educación, CNED) and the Ministry of

Education. The dataset includes a comprehensive list of programs offered by all higher

education institutions in each academic year. The program-level data contain information

on tuition fees, field of study, length, geographical location, and application requirements,

among others.

Third, we obtain measures of academic proficiency from the SIMCE (Sistema de Medicion

de Calidad de la Educacion, Education Quality Measurement System), a yearly national test

that is part of an information system established by the Chilean government to periodically

evaluate learning outcomes across the country. The SIMCE tests are taken by elementary

(2nd, 4th, and 6th grades), middle school (8th grade), and high school (10th and 11th

grades) students. Their main goal is to provide information about the learning achievements

of students in a wide range of knowledge areas that are part of the national curriculum.

Currently, the tests measure numeracy and language and communication skills, as well as

knowledge of natural sciences, history, geography, social sciences, and English. Since its

introduction (1980s) the SIMCE has undergone changes regarding the subjects tested as well

as the grades in which the examinations are administered. Each year, the National Agency

for the Quality of Education decides which grades will take the SIMCE tests as well as the

corresponding areas of knowledge that will be evaluated. It is common for students to take

15

the SIMCE two or three times during their school years. We choose to focus on the 4th grade

SIMCE, as it is the first SIMCE the class of 2016 took and was also taken by the previous

cohort of students.

4 Empirical Strategy

The college choice literature has systematically found that students’ preferences are sensitive

to tuition fees and that changes in relative prices may affect the way students apply to and

enroll in different colleges and programs (for evidence from Chile, see Alvarado et al., 2012;

Hastings et al., 2016; Espinoza, 2017; Bucarey, 2018; Solis, 2017). The tuition-free policy

changed the relative price of different majors for a subpopulation of prospective students.

The value that student i gives program j is a function of price (tuition) pi,j and the expected

lifetime earnings that she may obtain with such a degree. If student i qualifies for the tuition

free program, then pi,j = 0 for any j offered by a participating university. Therefore, the

value derived from the different programs j will not be the same with and without the policy,

and student application behavior will change accordingly.

As Table 2 shows, teaching degree programs have the lowest tuition fees in Chile. Equalizing

their price to more expensive majors in other fields may affect the pool of students applying

and pursuing a teaching programs. Lowering to zero the tuition fees charged by programs

in other fields may have made them more attractive to students who would have chosen

teaching programs in the absence of the free-college policy, and whose preferences for a

teaching degree may have been influenced by the relatively low price. Therefore, tuition-free

college has the potential to change students’ choices by equalizing the cost of pursuing more

expensive degrees in other fields with the cost of attending teaching programs. The impact of

such change on the pool of students pursuing a degree and the extent to which students may

switch from teaching majors to majors in other fields is the empirical question we address in

16

this paper.

We use the introduction of tuition-free college as a natural experiment to analyze the sorting

of students into majors. Especially, how it alters the pool of applicants who choose to pursue

teaching programs, given that such a pool had already been responding to the incentives

provided by the BVP scholarship. As explained above, eligibility for the BVP scholarship

required scoring more than 600 points in the college entrance exam PSU. Hence, the base of

our empirical strategy throughout the paper will entail the comparison of application behavior

before and after the introduction of tuition-free college among those who score above and

below 600 points in the PSU. In that sense, ours is a difference-in-difference strategy, where

one difference comes from the introduction of the policy and the second difference comes

from the BVP scholarship eligibility criterium.12

Our typical estimation strategy can be described by the following:

Pi,t = βPostt + γ1[PSUi,t ≥ 600] + δPostt1[PSUi,t ≥ 600] +Xi,tθ + µs + εi,t (1)

where Pi,t is the outcome of interest (e.g., applying to a teaching program or not) of student i

at time t, Postt takes the value of 1 if t ≥ 2016 the year of the introduction of the tuition-free

college and 0 otherwise, 1[PSUi,t ≥ 600] is an indicator variable that takes the value of 1 when

student i scores more that 600 points in the college entrance exam, Xi,t are individual level

controls like gender and contains a constant, and µs is a school-level or comuna-level fixed-

effect. Hence, the parameter δ corresponds to the effect that the introduction of tuition-free

college had among those that would have been eligible for the BVP scholarship.13

12We assume that the distribution of college preferences and income are equal in the 2015 and 2016 highschool cohorts. Nonetheless, our results are robust if we take into account that the free-tuition policy couldhave brought "new entrants" into higher education, as we show in Appendix G. We also need to assume thatthe effect of the BVP is stable across years, which Castro-Zarzur (2018) shows in her paper.

13Despite the existence of other scholarships applicable to any field of study, in 2015, the BVP grant wasthe only tuition scholarship covering full tuition costs. In particular, while the BBIC and the BJGM wereavailable for students coming from income deciles 1-7 and did not limit the field of study, they only covered upto 80% of the full tuition costs. This 20% difference corresponds to approximately 150% of the legal minimumwage, which is an economically significant figure if we take into account the per-capita income bounds of the

17

We extend strategy (1) to incorporate the fact that college was made tuition free for only

those who come from the bottom half of the income distribution. Incidentally, that subgroup

is the one that has responded the most to the incentives provided by the BVP scholarship

(Castro-Zarzur, 2018). To do so, we use Y gi,t as an indicator variable that takes the value of 1

if student i’s income belongs to category g and 0 otherwise. By convention, we use Y 0 = 1 to

account for the base terms that have no income-group interaction (i.e., β0Postt+γ01[PSUi,t ≥

600] + δgPostt1[PSUi,t ≥ 600]). Thus we can write our estimating equation as:

Pi,t =G−1∑g=0

(βgY

gi,tPostt + γgY

gi,t1[PSUi,t ≥ 600] + δgY

gi,tPostt1[PSUi,t ≥ 600]

)+Xi,tθ + µs + εi,t (2)

By providing an estimate of δg for each income group, we can see if the policy affected more

students from poorer families than wealthier ones, as we anticipate.

We proxy Y g with two measures that are known to closely correlate with family income: the

education level of the student’s mother and the type of high school in which the student

finished. We do so because we do not perfectly observe per-capita family income at the time

of application. But also, by proxying income with indicators of long term socio-economic

status, we isolate our estimates from possible strategic behaviors in which families at the

margin could reduce their labor supply in order to become eligible for the big financial

relief of not having to pay college tuition. In particular, we split mother education into

four categories: high school dropouts (19.4%), high school graduates (36.9%), some tertiary

education (22.4%), and college graduates (21.3%). Regarding school type, we use the fact

that primary and secondary schools in Chile are classified in three main categories: private,

public, and voucher. School types exhibit significant differences between them and are highly

income deciles used in the 2015 scholarship assignment process. For instance, per-capita incomes in the fifthdecile ranged between 56% and 69% of the legal minimum wage (see Table A1 in the Appendix). Therefore,the BVP was effectively enabling an important portion of low-income students’ access to higher education.Changes in the application behavior of precisely this population of students allows us to identify δ.

18

Figure 2: Recipients of Tuition Scholarships in 2016

(a) By Type of School

62.7

20.9

16.4

48.4

24.7

26.9

6.7

9.6

83.7

0

20

40

60

80

100

%

Public Voucher Private

Recipients of Tuition Scholarships by Type of School (2016)

Free College Other No Scholarship

(b) By Mother’s Education

70.8

17.5

11.7

56.2

22.8

21.0

32.1

23.0

44.9

14.5

17.7

67.8

0

20

40

60

80

100

%

S<12 S=12 12<S<16 S>=16

Recipients of Tuition Scholarships by Mother’s Education (2016)

Free College Other No Scholarship

Note: Both figures show the share of students from the 2016 graduating cohort who apply to college andobtain (or not) tuition scholarships. The left panel shows those fractions by school type. The right paneldoes so by mother’s education. S < 12: high school dropout, S = 12: high school graduate, 12 < S < 16:technical education or incomplete college, and S >= 16: complete college or above.

correlated with students’ socioeconomic background (Correa et al., 2014). Public schools are

run by municipalities and publicly funded. Voucher schools, which can be for-profit or not-

for-profit, are privately owned and receive a per-student subsidy (voucher) from the state.14

Finally private schools do not receive any public funds. Wealthier students typically attend

private schools, voucher schools are highly demanded by the middle class, and poorer students

typically attend public schools (Elacqua and Santos, 2013; Sánchez, 2018).

Figure 2 shows that school type and mother’s education are good predictors of access to need-

based college scholarships, thus reflecting the incidence of long-term credit constraints.15 For

instance, while 83.6% of applicants coming from public schools have access to a scholarship,14The fraction of the cost covered by the voucher varies from family to family depending on its socio-

economic status and the monthly cost of the school.15Tables A2 and A3 in Appendix C present descriptive evidence further supporting the use of mother’s

education and school type as proxies of family income.

19

only 16.3% of applicants coming from private schools do so. That is almost an exact reversal in

the likelihood of need-based accessing scholarships. In that domain, we see that the incidence

of access to scholarships among applicants graduating from voucher schools is closer to that of

public schools than to that of private schools. Figure 2b shows that 88.3% of applicants whose

mothers were high school dropouts required some kind of scholarship, while less than a third

of students whose mothers completed a college degree applied to college with a scholarship

at hand.

We are interested in evaluating the effect of the tuition-free college policy not only on the

application behavior and preferences between majors of the incoming students, but also on

how that sorting changed the distribution of academic quality of the students across majors.

This stems from the extensive evidence showing that high-achieving students are more likely

to become better teachers (Auguste et al., 2010; Seng Tang, 2015). To do so, we implement

the following specification that relies on the introduction of the policy and its differential

effects across income groups due to its eligibility criterium:

Scorei,t =G−1∑g=0

(βgY

gi,tPosttTeachi,t

)+Xi,tθ + µs + εi,t (3)

where Teachi,t is an indicator variable that takes the value of 1 if student i lists a teaching

program as one of her choices when applying to college. That way, we will compare the

change in the average academic proficiency after the introduction of free college among those

who pursue a teaching degree and those who do not, at different levels of family income.

As measures of academic proficiency, we use scores from SIMCE standardized tests measured

when the students were in 4th grade of elementary school, around 10 years old. that is, eight

years before they consider their tertiary education decision. We refer to those past scores

instead of using the PSU for this particular estimation because application decisions may

depend on the PSU score obtained. Therefore, there is a feedback process between scores

and application behavior that would bias the results. Instead, the 4th grade SIMCE scores,

20

Table 4: Change in Application Behavior by PSU range

Panel A: PSU<600Before After Diff. (%) p-value

Business 0.102 0.108 6.05 0.001Education 0.111 0.109 -1.85 0.293Health 0.288 0.299 3.54 0.000Social Sciences/Humanities 0.139 0.145 3.83 0.015STEM 0.220 0.199 -9.44 0.000Others 0.140 0.141 0.77 0.621

Panel B: PSU≥600Before After Diff. (%) p-value

Business 0.105 0.105 0.68 0.792Education 0.044 0.036 -17.78 0.000Health 0.249 0.279 12.13 0.000Social Sciences/Humanities 0.105 0.108 2.54 0.325STEM 0.350 0.319 -8.69 0.000Others 0.148 0.153 3.10 0.144Note: For students applying to degrees in the centralized matching system, we comparethe probability of applying to a degree (in each of the fields) as top choice, before andafter the implementation of the tuition-free policy. We restrict the sample to studentswho graduated from high school the year before entering higher education. The lasttwo columns test the statistical significance of these differences.

having no bearing on the college acceptance decisions, do not influence students’ application

behavior.

5 Results

5.1 Application Behavior

We first show that the introduction of free college had an immediate effect on student appli-

cation behavior. Table 4 compares the field of study of the students’ most preferred choice

before and after free college. It presents simple mean comparisons between application fre-

quencies to different fields. We split the sample in two. The top panel shows the frequency of

applications among students scoring below the 600-point threshold that defines the eligibility

21

to the BVP scholarship. Panel B replicates these statistics for students who are eligible for

the BVP, those who score above 600 points. The last two columns show the before-after

difference and its statistical significance. The table shows a statistically significant drop in

the fraction of students applying to teaching degrees as their most preferred choice among

top performers. In the year before the introduction of free college, 4.4% of students scoring

more than 600 points applied to a teaching degree as their most preferred choice. After the

policy the fraction dropped to 3.6%. That is, among those that are eligible for the BVP

scholarship, the introduction of free college causes a decline in the probability of applying a

teaching degree as top choice of about 17.8%. The drop is largest across all fields of study.

Such behavior supports the hypothesis that the introduction of free college decreased the re-

turn of pursuing a teaching degree vis-a-vis degrees in other degrees that compete for similar

students. In contrast, such behavior is not mirrored by students scoring below the 600-point

threshold. There is only a slight and non-significant decline from 11.1% to 10.9%.16

Next, we use the regression framework detailed in Section 4 to estimate the effect of free

college on student application behavior. We estimate the model using two alternative de-

pendent variables. The first takes the value of 1 if the student applied to a teaching degree,

independent of the order in which she listed it (Teaching as a choice). In the second one,

the dependent variable takes the value of 1 if the student applied to a teaching degree as her

top choice (Teaching as top choice). As indicated in Section 4, we do not use precise house-

holds income data to exactly pin down each student’s eligibility for free college. Instead, we

proxy household income with two variables that are well known to closely correlate with it:

type of school (i.e., public, voucher or private) and student mothers’ education. Results are

presented in Tables 5 and 6. We report the marginal effects for students scoring above and

below BVP scholarship’s eligibility criteria.16Table 4 shows that STEM majors also experienced a decline in the fraction of applicants. We analyze

this interesting phenomenon in a separate paper (Castro-Zarzur et al., 2018). Importantly for the purpose ofthe this paper, the drop in STEM applications is common to both sides of the 600-point threshold. Therefore,it is not a confounding factor in the margin we are interested in.

22

Table 5: Effect of Free College on Application Behavior to Teaching Majors by School Type

Teaching as a Choice Teaching as Top ChoiceCoef. Std.Err. Coef. Std.Err.

PublicPSU < 600 0.002 (0.006) -0.006 (0.004)PSU ≥ 600 -0.023*** (0.008) -0.016*** (0.004)Diff -0.025*** -0.010*

VoucherPSU < 600 -0.004 (0.003) -0.008*** (0.003)PSU ≥ 600 -0.014*** (0.005) -0.012*** (0.003)Diff -0.010* -0.004

PrivatePSU < 600 0.000 (0.007) 0.003 (0.005)PSU ≥ 600 -0.001 (0.003) -0.001 (0.002)Diff -0.001 -0.003

Obs. 154,277 154,277Avg. dep. variable 0.171 0.086

Note: We present the size of the effect for each category calculated based onthe regression results presented in Table A12 in the Appendix. All regressionsinclude comuna fixed-effects, gender, and linear PSU score controls. The 600threshold for the PSU score was chosen based on the minimum score requiredto apply for BVP scholarship. Standard errors clustered at the school level.*p < 0.1, **p < 0.05, ***p < 0.05.

Table 5 proxies household income with school type. Its first column shows that the probability

of listing a teaching major as a choice falls significantly only among those coming from public

and voucher schools that scored above the 600-point threshold. That is, with the introduction

of the free-college policy, the probability of considering a teaching major falls for relatively

poor BVP-eligible students. The drops are not only statistically significant but economically

meaningful. The likelihood of listing a teaching major among these students falls by 2.7

(public) and 1.5 (voucher) percentage points. These drops correspond to a 16% and 9%

decrease, respectively. Furthermore, these results are even larger among those who listed a

teaching major as their top choice. The likelihood for relatively poor BVP-eligible students

of doing so decreased by 18.6% and 14% for students from public and voucher schools,

respectively. In contrast, we find no significant changes in the probability of applying to

teaching programs among the relatively wealthy students (i.e., those graduating from private

23

schools) who are not eligible for free college.

Interestingly, we find a response, albeit smaller, on the number of students listing an education

major as their top choice among voucher school students scoring below the BVP eligibility

threshold. We argue that this decline results from the fact that some voucher school students

are eligible for free college, and some of them, who now consider other majors as their top

choice, would have found in teaching the only pathway to a college degree when college was

not free for them, as teaching majors were usually the cheapest ones.

Our findings remain overall consistent when we proxy household income with mother’s years

of education in Table 6.17 BVP-eligible students from relatively poor socioeconomic back-

grounds (i.e. with relatively less educated mothers) respond to the introduction of free college

by opting out of the teaching profession. The first column of Table 6 shows that the probabil-

ity of applying to a teaching program decreases significantly for BVP-eligible students with

high school dropout mothers (18%), high school graduate mothers (14.3%), and to some

extent to those whose mothers had technical degrees or incomplete college who decreased

their teaching major applications by 10%. Additionally, in line with our previous finding, the

probability of applying to a teaching program as top choice drops for poorer students who

score below the 600-point threshold, possibly reflecting the substitution of a relatively cheap

major with others that they can now afford due to the free-college policy.

The introduction of free college increased the number of applications by 3.4 percentage points,

mostly coming from outside the top three deciles of the PSU distribution (Castro-Zarzur et al.,

2018). Thus, some of the effect found so far could be due, in part, to “new entrants”—people

who would not have applied in the absence of the policy—rather than to a “reshuffle” across17Results in Tables 5 and 6 are not strictly comparable because the ones collected in the former use comuna-

level fixed-effects, while the ones presented in the latter come from estimations using school-level fixed-effects.The use of school-level fixed-effects may be more desirable as they capture any unobserved school traits thatmay correlate with college application. However, being the school type a time invariant school characteristics,its effect on a given outcome is not identified in a school-level fixed-effects specification. We consider thatthe results coming from the models with school types provide relevant insights and thus are worth reportingdespite their lack of school-level fixed-effects.

24

Table 6: Effect of Free College on Application Behavior to Teaching Majors by Mother’sSchooling

Teaching as a Choice Teaching as Top ChoiceCoef. Std.Err. Coef. Std.Err.

S < 12 (19.4%)PSU < 600 -0.005 (0.006) -0.016*** (0.004)PSU ≥ 600 -0.027** (0.012) -0.011 (0.007)Diff -0.022* 0.005

S = 12 (36.9%)PSU < 600 -0.004 (0.004) -0.006* (0.003)PSU ≥ 600 -0.019*** (0.006) -0.016*** (0.004)Diff -0.014** -0.010*

12 < S < 16 (22.4%)PSU < 600 -0.003 (0.006) -0.004 (0.004)PSU ≥ 600 -0.010* (0.005) -0.008** (0.004)Diff -0.007 -0.004

S ≥ 16 (21.3%)PSU < 600 0.005 (0.007) 0.002 (0.005)PSU ≥ 600 -0.007* (0.004) -0.004 (0.002)Diff -0.012 -0.006

Obs. 140,984 140,984Avg. dep. variable 0.173 0.086

Note: We present the size of the effect for each category calculated based on theregression results presented in Table A13 in the Appendix. All regressions includeschool fixed-effects, gender, and linear PSU score controls. The 600 threshold forthe PSU score was chosen based on the minimum score required to apply for BVPscholarship. Standard errors clustered at the school level. *p < 0.1, **p < 0.05,***p < 0.05.

majors of people who would have applied anyway. To address that, we rerun our estimations

on a 2016 subsample that is most likely to have applied without the implementation of free

college. We define that subsample based on a propensity score estimated from observed

characteristics of students and their application decisions in 2015. That is, we model college

application decisions with gender, mother’s education, 4th grade math and language SIMCE

scores and the school students come from. The model reports an R2 of 0.91; details can be

found in Appendix G. We take the estimated parameters and predict the propensity of having

applied to college based on their observable characteristics for the 2016 sample. Thus, we

drop from the 2016 sample those who applied but were less likely to do so according to the

25

propensity score. We drop around 9,000 applicants from 2016 in order to match the number

of applicants in 2015.

The results of the estimations on the subsamples without “new entrants” are reported in

Tables A8 and A9 in Appendix G. They are very similar to the ones obtained using the

whole sample, indicating that most, if not all, the effects on applications to education majors

found so far are due to changes in major preferences of students who would have applied to

college even if the tuition-free policy had not been implemented.

Changes in application behavior due to the introduction of free college may be evident in

the extensive margin (i.e., whether the student applied to a teaching program or not) but in

the intensive margin as well. To explore that, we exploit the information contained in the

application list submitted to the clearinghouse, in which students rank in order of desire the

programs (i.e., university-major pairs) they consider pursuing.

Tables 7 and 8 present changes in the probability of applying to a teaching major before and

after the introduction of the free college reform for four different ranking positions. These

tables confirm our previous findings: first, the observed drops in applications to education

majors are driven by the application behavior changes among relatively poor BVP-eligible

students; and second, these drops are driven by students no longer listing educations majors as

their top choice. These declines are consistent with the fact that financially constrained, high-

performing individuals who before the free-tuition reform could only access higher education

with the BVP grant can now choose from a wide range of majors besides teaching, many of

which have higher expected returns. Indeed, the introduction of the free-college policy not

only resulted in less application to education majors from academically better individuals,

but also in a deterioration of the relative attractiveness of the teaching career,even for those

top-performing students who consider it as an option.

Our results also shows that for students with PSU scores below 600, the probability of

choosing a teaching major as their fourth or higher most preferred choice increased by 8.4%.

26

Table 7: Preference Order of Education Programs Before and After Free College (by SchoolType)

PSU < 600 PSU ≥ 600Education as Before After Dif.(%) Before After Dif.(%)

Panel A: PublicFirst choice 0.123 0.122 -0.908 0.059 0.044 -24.088**Second choice 0.033 0.032 -0.566 0.014 0.015 2.681Third choice 0.028 0.029 5.573 0.018 0.020 12.615≥ Fourth choice 0.054 0.059 10.728** 0.053 0.045 -14.910*

Panel B: VoucherFirst choice 0.113 0.108 -4.201* 0.056 0.044 -21.766***Second choice 0.027 0.026 -6.292 0.015 0.013 -11.076Third choice 0.023 0.024 3.288 0.016 0.015 -5.398≥ Fourth choice 0.045 0.048 8.829** 0.038 0.038 -0.697

Panel C: PrivateFirst choice 0.067 0.070 4.690 0.021 0.021 -0.067Second choice 0.015 0.017 17.792 0.006 0.006 7.479Third choice 0.017 0.016 -6.866 0.010 0.009 -6.020≥ Fourth choice 0.032 0.028 -12.616 0.018 0.018 -2.364

Note: The average student lists five programs. In case the student listed more than one teachingprogram among her preferences, we chose for the regressions the minimum rank among them. The600-point threshold for the PSU score was chosen based on the minimum score required to applyfor BVP scholarship. *p < 0.1, **p < 0.05, ***p < 0.05.

Likewise, such an increase is limited to relatively poor individuals (i.e., those coming from

public and voucher schools or less-educated mothers), as reported in Tables 7 and 8.

Interestingly, Table 8 shows that students with PSU scores below 600 who have high school

dropout mothers decreased their likelihood of choosing teaching as their top choice and, at

the same time, increased their likelihood of listing teaching as their third choice or as their

fourth of higher choice. These results suggest a reordering of the preference for teaching

among relatively low-performing students from the poorest backgrounds, by which education

majors lose their appeal and are downgraded to less-preferred options. This is again consistent

with the fact that free college allows the relatively low performers to switch away from the

comparatively cheaper teaching programs to majors that are costlier and generally tend to

have higher expected returns.

27

Table 8: Preference Order of Education Programs Before and After Free College (by Mother’sEducation)

PSU < 600 PSU ≥ 600Education as Before After Dif.(%) Before After Dif.(%)

Panel A: S < 12First choice 0.130 0.119 -8.749** 0.061 0.051 -15.342Second choice 0.032 0.031 -2.059 0.018 0.017 -1.865Third choice 0.024 0.028 20.266** 0.021 0.021 3.342≥ Fourth choice 0.047 0.055 17.781** 0.050 0.037 -26.770**

Panel B: S = 12First choice 0.117 0.115 -1.758 0.060 0.046 -23.242***Second choice 0.027 0.028 0.532 0.014 0.013 -10.189Third choice 0.025 0.025 -0.773 0.016 0.016 -0.834≥ Fourth choice 0.049 0.051 3.892 0.042 0.042 -0.518

Panel C: 12 < S < 16First choice 0.099 0.099 -0.554 0.045 0.036 -18.679**Second choice 0.025 0.025 -3.081 0.013 0.010 -16.407Third choice 0.023 0.024 4.560 0.013 0.014 4.787≥ Fourth choice 0.042 0.044 4.101 0.030 0.030 -0.807

Panel D: S ≥ 16First choice 0.079 0.081 3.580 0.028 0.026 -8.460Second choice 0.022 0.022 -0.307 0.008 0.009 18.235Third choice 0.021 0.019 -10.249 0.013 0.011 -11.441≥ Fourth choice 0.040 0.045 11.211 0.027 0.024 -8.582

Note: The average student lists five programs. In case the student listed more than one teachingprogram among her preferences, we chose for the regressions the minimum rank among them. The600-point threshold for the PSU score was chosen based on the minimum score required to apply tothe BVP scholarship. *p < 0.1, **p < 0.05, ***p < 0.05.

Hence, in conjunction with our previous results, we find that the introduction of free college

had effects in both the extensive and the intensive margin of teaching majors application

behavior. However, the effect is heterogeneous across different groups of students. Free

college reduced the probability of considering a teaching profession as an option among

those who would have been eligible for the BVP scholarship. In addition, for those who

did consider a teaching profession among their choices, the introduction of free college made

teaching majors less attractive: they were less likely to put them as their first choice and

more likely to list them as a latter, less preferred option.

28

Table 9: Effect of Free College on PSU Score of Accepted Applicants

Panel A: OverallN=110935 Other Education Diff.

Post -0.362 (0.482) -2.671*** (0.784) -2.309**

Panel B: School TypeN=110719 Other Education Diff.

Post#Public -2.710** (1.268) -4.724*** (1.429) -2.014Post#Voucher -0.104 (0.655) -2.709*** (1.018) -2.605**Post#Private 3.198*** (1.059) 2.789 (2.598) -0.409

Panel C: Mother’s EducationN=100814 Other Education Diff.

Post#S < 12 -2.524*** (0.960) -5.754*** (1.555) -3.230*Post#S=12 -1.389* (0.744) -3.689*** (1.201) -2.300Post#12 < S < 16 1.803** (0.902) -1.189 (1.644) -2.992Post#S ≥ 16 0.445 (0.916) 2.214 (2.082) 1.769

Note: All regressions include gender controls. Regressions in Panel A and C include school fixed-effects. Regressions in Panel B include comuna fixed-effects. Column Diff presents the differencebetween the point estimates of Education and Other and tests for whether they are statisticallydifferent from each other or not. Standard errors clustered at the school level in parentheses.*p < 0.1, **p < 0.05, ***p < 0.05.

5.2 Effect on Enrollment

The change in application behavior due to the introduction of free college translated to a

change in the relative academic proficiency of the students that were offered admission in

different programs as measured by the PSU score. In Table 9, we present the changes in

PSU scores due to the introduction of free college for education majors and non-education

majors separately. In Panel A, we show that while the introduction of free college did not

change the average score of the students accepted to non-education majors, the scores among

the accepted to education ones did deteriorate. We find an overall drop of 2.67 percentage

points. That amounts to a decrease in the score of the new entrants to the teaching majors

of almost 4% of a standard deviation.18

Panels B and C of Table 9 include the interactions with income proxies. Panel B reports the18While the standard deviation of the PSU among the entire population that took the exam is 110, when

we limit the sample to those who ended up being accepted in traditional universities, the standard deviationof the score becomes 68.5.

29

results by school type, while Panel C does so by mother’s education level. Both cases provide

the same insights. The deterioration in the quality of the students admitted is limited to those

relatively poor. The average score among those admitted to education programs that came

from the poorest backgrounds fell by 9.1% of a standard deviation. Interestingly, while the

quality of those from voucher schools and high school graduate mothers that were accepted in

non-education majors remains mostly unchanged, the performance of those who were offered

admission to education majors after the implementation of free college is significantly lower.19

These findings provide evidence in favor of a mechanism in which the BVP scholarship had

brought relatively high-quality students who could not afford college. Once the relative

profitability of the teaching programs changed again due to the free tuition policy that

covered the other majors, that subpopulation of high-scoring low-income students shifted to

other disciplines, as they no longer saw the teaching majors as the only pathway to a college

degree.

In Table 10, we further explore the deterioration of the academic proficiency of applicants to

the teaching majors caused by the introduction of the tuition-free college policy. We regress

measures of academic proficiency on a dummy indicating the introduction of the policy in

samples that consist of those who listed the education major within their 10 preferred degrees

(Any) and those who did so as their top choice (Top). As indicated in the empirical strategy,

we use the math and language SIMCE scores measured when the students were in 4th grade

of elementary school as measures of academic proficiency, that is, eight years before they

consider their tertiary education decision. This way we avoid the feedback process between

PSU scores and application behavior that would bias the results.

We find that the introduction of free college changed the pool of applicants to teaching majors

in such a way that it ended up being comprised by students who were on average 14% of a19In Table A11 in the Appendix, we show the estimations using data on enrolled students instead of

accepted students. Our findings remain the same, albeit sightly smaller point estimates. Differences are notstatistically significant. In Appendix H, on the other hand, we perform the same estimations using mathscores in 4th grade of elementary school. The results hold.

30

Table10

:Effe

ctof

Free

College

onSIMCE

Scores

ofApp

lican

tsto

Edu

cation

Program

s

Overall

Scho

olTyp

eMother’sEdu

cation

Any

Top

Any

Top

Any

Top

Pan

elA:Math

Pos

t-6.259

***

-7.112

***

Pos

t#Pub

l-9.459

***

-8.082

***

Pos

t#S<

12-7.469

***

-6.736

***

(0.664

)(0.725

)(1.157

)(1.235

)(1.419

)(1.560

)Pos

t#Vou

-7.637

***

-9.029

***

Pos

t#S=

12-7.870

***

-8.639

***

(0.827

)(0.896

)(1.080

)(1.153

)Pos

t#Pri

v-0.933

-2.132

Pos

t#S16

12

-4.803

***

-7.277

***

(1.987

)(2.293

)(1.444

)(1.652

)Pos

t#S≥

16-1.272

-2.364

(1.747

)(1.903

)

Obs.

1146

7951

031

1143

6850

897

1058

1947

079

Pan

elB:Lan

guage

Pos

t-3.810

***

-4.439

***

Pos

t#Pub

l-5.629

***

-5.168

***

Pos

t#S<

12-6.855

***

-5.915

***

(0.654

)(0.706

)(1.114

)(1.235

)(1.433

)(1.576

)Pos

t#Vou

-5.496

***

-6.557

***

Pos

t#S=

12-5.291

***

-6.088

***

(0.796

)(0.851

)(1.058

)(1.127

)Pos

t#Pri

v0.72

1-0.991

Pos

t#S16

12

-1.287

-2.270

(2.188

)(2.404

)(1.512

)(1.613

)Pos

t#S≥

161.60

00.709

(1.736

)(1.912

)

Obs.

1144

5450

953

1141

4650

819

1056

3347

021

Note:

Allregression

sinclud

egend

ercontrols.Allregression

sinclud

escho

olfix

ed-effe

cts,

except

forscho

oltype

regression

s,which

includ

ecomun

afix

ed-effe

cts.

Stan

dard

deviationof

thescores

inthesubsam

ples

areas

follo

ws.

Math(A

ny):

44.45,

Math(T

op):

45.24,

Lang

uage

(Any

):43.31,

andLa

ngua

ge(T

op):

43.77.

Stan

dard

errors

clusteredat

thescho

ollevelin

parentheses.

*p<

0.1,

**p<

0.05

,***p<

0.05

.

31

standard deviation less proficient in math and 8.8% of a standard deviation less proficient

in language, as measured by their 4th grade scores. As expected, we see a stronger effect

among relatively poor students. In fact, we observe drops of 20.5% of a standard deviation

in the math scores and 12.8% of a standard deviation in the language scores among students

from public schools, and 16.24% of a standard deviation in the math scores and 15.8% of a

standard deviation in the language scores among students whose mothers were high school

dropouts. These figures contrast with those reported for students from affluent backgrounds

in which no statistically significant effects were found. All these results provide evidence

in favor of the argument, indicating that the large declines of the scores of applicants to

teaching majors is the result of two forces: the introduction of relatively poor low-scoring

students to the tertiary education system as college became affordable, and a substitution

effect among relatively poor high-scoring students who leave the education majors for other

disciplines, as the BVP scholarship is no longer giving education majors an advantage in

terms of profitability.

6 Beyond the Difference-in-Difference

We further use BVP’s score eligibility cutoff (i.e., PSU ≥ 600) to estimate the effects of the

introduction of free college on applications to teaching majors. To that effect, we compare

the differences in application behavior before and after the introduction of free college. We

unpack our difference-in-difference results by identifying heterogeneous effects in application

behavior to teaching majors at different PSU levels. We estimate non-parametric functions

of PSU scores on application behavior to the left and to the right of the PSU = 600 cutoff

and before and after the introduction of the free-college policy.

Figures 3 and 4 present our results. In each figure, we present the non-parametric results by

each type of school. Figure 3 considers the probability of listing a teaching major anywhere

32

in the preference list, while Figure 4 considers the probability of listing a teaching major as

their top choice. As expected, we consistently find no differences in application behavior for

those whose PSU scores were less than 600. They simply had the same incentives before

and after the introduction of free college as they were never eligible for the BVP scholarship.

The interesting results come from the right of the cutoff. First, in line with our parametric

difference-in-difference results, we find significant differences in the application behavior of

students in public and voucher schools who scored more than 600 and were or would have

been eligible for the BVP scholarship. We do not find those gaps among students coming

from private schools. Second, our findings suggest that by 2015 the BVP was not adding

(bringing in) extra students to the teaching profession, but avoiding larger shortages of

qualified applicants to teaching majors.20 With the introduction of the tuition-free policy,

those shortages deepened as it made the BVP scholarship innocuous. Third, Figures 3 and

4 show that the gaps in application behavior do not open right at the PSU = 600 BVP

eligibility cutoff. Instead, they open 5 to 10 PSU points to the right of that threshold. This

may seem puzzling, since the elimination of the BVP incentive should in principle affect all

those that would have been eligible. However as we have established, the introduction of free

college made teaching majors less attractive to BVP-eligible high school graduates relative

to other majors. Then, the reference score for those students is not PSU = 600 but the

scores required to get into the majors in which they intend to enroll. The vertical lines in

Figures 3 and 4 represent those scores, namely, the average scores of applicants accepted

to health, social sciences and humanities, and STEM programs in 2015.21 Those scores are

to the right of the PSU = 600 threshold and in the case of health, social sciences and

humanities majors—which were the ones with the greatest inflow of PSU > 600 students

after the introduction of the free tuition policy—overlay remarkably well with the PSU scores20This is consistent with findings reported in Castro-Zarzur (2018) where, using a regression discontinuity

approach, finds that the positive impacts of the BVP diminished with each incoming cohort.21The scores of accepted applicants from past cohorts to each program are publicly available in

www.psu.demre.cl and in specific university websites like www.admision.uai.cl and http://www.admisionyregistros.uc.cl. They are made public intentionally year after year in order to inform prospec-tive applicants about the scores they require to maximize their chances of getting into a given program.

33

Figure 3: Probability of Listing a Teaching Major Within Choice Set and PSU Scoreby School Type

Health

SocialSc.

& Human

STEM

.1.1

5.2

.25

550 560 570 580 590 600 610 620 630 640 650

Pre Post

Public

Health

SocialSc.

& Human

STEM

.1.1

5.2

.25

550 560 570 580 590 600 610 620 630 640 650

Pre Post

Voucher

Health

SocialSc.

& Human

STEM

.05

.1.1

5.2

550 560 570 580 590 600 610 620 630 640 650

Pre Post

Private

Note: Local mean smoothing estimated with Epanechnikov kernel using optimal bandwiths. Dependentvariable takes the value of 1 of student listed a teaching major anywhere in her choice set and 0 otherwise.Capped spikes represent confidence intervals at the 90% level. Vertical lines present the average PSU scoreof those accepted to health, social sciences and humanities, and STEM majors in 2015.

at which the application behavior gaps begin to emerge.

7 Discussion

In this paper, we study the extent to which making college tuition free affects the pool of

students pursuing a teaching career. We exploit a major reform carried out in Chile in 2016

that eliminated tuition costs of undergraduate studies for students from the lower half of

the income distribution. We study how the introduction of a free-college policy affects the

34

Figure 4: Probability of Listing a Teaching Major as Top Choice and PSU Scoreby School Type

Health

SocialSc.

& Human

STEM

.02

.04

.06

.08

.1.1

2

550 560 570 580 590 600 610 620 630 640 650

Pre Post

Public

Health

SocialSc.

& Human

STEM

.04

.06

.08

.1.1

2

550 560 570 580 590 600 610 620 630 640 650

Pre Post

Voucher

Health

SocialSc.

& Human

STEM0.0

2.0

4.0

6.0

8

550 560 570 580 590 600 610 620 630 640 650

Pre Post

Private

Note: Local mean smoothing estimated with Epanechnikov kernel using optimal bandwiths. Dependentvariable takes the value of 1 of student listed a teaching major as her top choice and 0 otherwise. Cappedspikes represent confidence intervals at the 90% level. Vertical lines present the average PSU score of thoseaccepted to health, social sciences and humanities, and STEM majors in 2015.

35

application behavior of students who were eligible for the BVP tuition grant, a scholarship

implemented in 2011 that had been successful in bringing high-quality students into teaching

majors. We do so because the free-college policy equalizes the relative price of studying a

wide range of different majors and therefore it has the potential of offsetting the incentives

set out by the BVP.

Our results suggest that granting tuition-free access to college for individuals coming from

the lower half of the income distribution decreased the demand for teaching programs of

top-performing students. In consequence, it deteriorated the academic quality of students

pursuing a teaching career. The introduction of the free-college policy reduced by about 17%

the probability of applying to a teaching major among the high-performing students (i.e.

BVP-eligible) that come from relatively poor backgrounds. In addition, among those who

still consider the teaching profession, the introduction of free college made teaching majors

less appealing as they are less likely to put them as their first choice and more likely to list

them as a less preferred option. Importantly, these changes in application behavior translated

into a deterioration of the academic qualifications of the pool of students who were offered

admission to teaching majors. While the introduction of free college did not change the

average PSU score of students accepted to non-education majors, the mean score of those

admitted to teaching programs fell by 4% of a standard deviation. Such decrease is explained

by a decrease in the scores of student from public or voucher schools and with relatively less

educated mothers, while the average PSU score of the relatively wealthier (students coming

from private schools) teaching freshmen remains unchanged.

Our findings are important for several reasons. First, they highlight the potential unintended

consequences of policies distorting equilibrium prices in markets such as higher education.

Second, our findings illustrates the lack of complementarity between two education policies,

the BVP and the free college reform. The latter offsets the benefits of the former. In our

case, our results suggest that with the introduction of free college, the achievements of the

36

BVP in terms of bringing high performing students into teaching majors were reversed. This

is worrisome as there is convincing evidence in favor of the notion that top performing high

school graduates are very likely to be better teachers and, in turn, high-quality teachers

positively impact the long-term outcomes of their students. Third, our findings also provide

an important input to the ongoing international debate on free college. For example, college

tuition-free policy has emerged as one of the central themes in recent U.S. Presidential Elec-

tions where the initiative has show to be appealing to a numbers of policymakers.22 This

paper shows a potential unintended consequence of such policy acting through a relative price

change which affects the sorting of students into majors. Our paper shows that in regards

to students’ preferences over teaching degrees, the free college policy has the potential to

negatively affect long-term teacher quality.

Finally, our findings are likely to apply even in the absence of previous ability-based schol-

arships, such as the BVP. Teaching majors tend to be the least costly in many countries,

including the United States where differential tuitions are being increasingly used; hence, in

the absence of tuition-free college the relatively cheaper teaching majors are a pathway to

higher education for low-income students. As free college equalizes the tuition cost across all

majors, this allows high performers from poorer backgrounds to substitute relatively cheap

majors with other - likely costlier - programs. In that sense, the BVP is just a vehicle for

identification as it allows us to implement a difference-in-difference strategy while providing

important insights on the potential impact of free college on self-selection into the teaching

profession in more general settings.

22Additionally, the free-college movement in the US has continuously gained strength since the last years.As of 2019, eleven states—Oregon, Nevada, Arkansas, New Jersey, Maryland, Tennessee, New York, Rhode Is-land, Delaware, Kentucky, and Indiana—have programs that typically offer two years of free tuition in certaincolleges for low- and middle-income students. More recently, in September 2019, New Mexico announced aplan for free college for state residents regardless of family income. Some examples of widely-known tuition-free college programs include the Kalamazoo Promise Program in Michigan and the State of New York’sExcelsior Scholarship program.

37

ReferencesAbraham, K. G. and Clark, M. A. (2006). Financial Aid and Students’ College Decisions Evidence

from the District of Columbia Tuition Assistance Grant Program. Journal of Human Resources,XLI(3):578–610. 2.3

Alfonso, M., Santiago, A., and Bassi, M. (2010). Estimating the Impact of Placing Top UniversityGraduates in Vulnerable Schools in Chile. Education Division SCLEDU Technical Notes, pages1–51. 2

Alvarado, M., Duarte, F., and Neilson, C. (2012). Efectos Preliminares de la Beca Vocación deProfesor. Technical report, Santiago de Chile. 2.2, 4

Araujo, M. C., Carneiro, P., Cruz-Aguayo, Y., and Schady, N. (2016). Teacher Quality and LearningOutcomes in Kindergarten. The Quarterly Journal of Economics, 131(3):1415–1453. 1

Auguste, B. G., Kihn, P., and Miller, M. (2010). Closing the Talent Gap: Attracting and RetainingTop-third Graduates to Careers in Teaching. McKinsey and Company. 2, 4

Balcázar, C. F. and Ñopo, H. (2014). Broken Gears: The Value Added of Higher Education onTeachers’ Academic Achievement. IZA Discussion Paper Series, pages 1–28. 1, 2.2

Bau, N. and Das, J. (2019). Teacher Value-Added in a Low-Income Country. American EconomicJournal: Applied . . . , pages 1–44. 1

Beyer, H., Hastings, J., Neilson, C., and Zimmerman, S. (2015). Connecting Student Loans to LaborMarket Outcomes: Policy Lessons from Chile. American Economic Review, 105(5):508–513. 8

Bonomelli, F. (2017). Seguimiento de la Beca Vocación de Profesor: Desde su implementaciónhasta puntos de encuentro con la Gratuidad y el Nuevo Sistema de Desarrollo Docente. TechnicalReport 3. 2.2

Bruns, B. and Luque, J. (2015). Great Teachers. How to raise stundent learning in Latin Americaand the Caribbean. World Bank Group, Washington, DC. 1

Bucarey, A. (2018). Who Pays for Free College? Crowding Out on Campus. 1, 4

Cabezas, V. and Claro, F. (2011). Valoración social del profesor en Chile: ¿cómo atraer a alumnostalentosos a estudiar pedagogía? Temas de la Agencia Publica, 6(42):1–18. 2.2

Castro-Zarzur, R. (2018). Can Service Scholarships be Effective in Bringing High-Quality Studentsto Teaching Programs in a Context of Increasing College Aid? Evidence from Chile. 1, 2.2, 12,4, 20

Castro-Zarzur, R., Espinoza, R., Sarzosa, M., and Sarz (2018). College Tuition and the Taste forSTEM Majors. pages 1–30. 16, 5.1

Centro de Estudios MINEDUC (2017). Estadisticas de la Educación 2016. Ministerio de Educación,Santiago de Chile. 2.1

Chetty, R., Friedman, J. N., and Rockoff, J. E. (2014). Measuring the Impacts of Teachers II: TeacherValue-Added and Student Outcomes in Adulthood. American Economic Review, 104(9):2633–2679. 1

38

Claro, F., Paredes, R. D., Bennett, M., and Wilson, T. (2013). Incentivos para estudiar pedagogía:El caso de la Beca Vocación de Profesor. Estudios Publicos, 131:37–59. 3, 2.2

Correa, J. A., Parro, F., and Reyes, L. (2014). The Effects of Vouchers on School Results: Evidencefrom Chile’s Targeted Voucher Program. Journal of Human Capital, 8(4):351–398. 4

Dynarski, S. M. (2003). Does Aid Matter? Measuring the Effect of Student Aid on College Atten-dance and Completion. American Economic Review, 93(1):279–288. 2.3

Eide, E., Goldhaber, D., and Brewer, D. (2004). The Teacher Labour Market and Teacher Quality.Oxford Review of Economic Policy, 20(2):230–244. 1, 2.2

Elacqua, G., Hincapie, D., Vegas, E., and Alfonso, M. (2018). Profesión: Profesor en AméricaLatina ¿Por qué se perdió el prestigio docente y cómo recuperarlo? Inter-American DevelopmentBank, Washington D.C. 1, 6, 2.2

Elacqua, G. and Santos, H. (2013). Preferencias reveladas de los proveedores de educación privadaen Chile. Gestión y Política Púbica, 23(1):85–129. 4

Espinoza, R. (2017). Loans for College: Strategic Pricing and Externalities. 4

Espinoza, R., Lee, S., and Lopez, H. (2017). Endogenous Market Formation: Theory and Evidencefrom Chilean College Admissions. 2.1

Espinoza, R. and Urzua, S. (2015). Las Consecuencias Económicas de un Sistema de EducaciónSuperior Gratuito en Chile. pages 10–44. 2.3

Espinoza, R. and Urzua, S. (2016). The Economic Returns to Higher Education: Funding, coverageand quality in Latin America. 1, 2.1, 2.2

Glazerman, S., Mayer, D., and Decker, P. (2005). Alternative routes to teaching: The impacts ofTeach for America on student achievement and other outcomes. Journal of policy analysis andmanagement, 25(1):75–96. 2

Gobierno de Chile (2017). Metodología de arancel de Referencia 2018 para universidades acreditadas.pages 1–5. 11

Gomez, C., Cabezas, V., Madeiros, M. P., Martinez, F., and Crisostomo, B. (2019). AttractingStudents to the Teaching Career: How Public Policy Affects Enrolment into Teaching TrainingPrograms. 2.2

Gonzalez-Velosa, C., Rucci, G., Urzua, S., and Sarzosa, M. (2015). Returns to Higher Education inChile and Colombia. IDB Working Paper Series. 1, 2.2

Hanushek, E. A., Piopiunik, M., and Wiederhold, S. (2018). The Value of Smarter Teachers: In-ternational Evidence on Teacher Cognitive Skills and Student Performance. Journal of HumanResources, pages 0317–8619R1–75. 1

Hastings, J., Neilson, C., and Zimmerman, S. (2013). Are Some Degrees Worth More Than Others?Evidence from College Admission Cutoffs in Chile. 1

39

Hastings, J. S., Neilson, C. A., Ramirez, A., and Zimmerman, S. D. (2016). (Un)informed collegeand major choice: Evidence from linked survey and administrative data. Economics of EducationReview, pages 1–16. 4

Konstantopoulos, S. (2011). Teacher Effects in Early Grades: Evidence From a Randomized Study.Teachers College Record, 113(7):1541–1565. 1

Lang, K. and Palacios, M. D. (2018). The Determinants of Teachers’ Occupational Choice. NBERWorking Paper, pages 1–41. 1, 2.2

Mizala, A. and Ñopo, H. (2016). Measuring the relative pay of school teachers in Latin America1997–2007. International Journal of Educational Development, 47:20–32. 4, 2.2

OECD (2005). Teachers Matter: Attracting, Developing and Retaining Effective Teachers. Technicalreport. 1, 3, 2.2

OECD (2017a). Education at a Glance 2017. OECD Indicators. OECD, Paris. 4

OECD (2017b). Education in Chile. Reviews of National Policies for Education, OECD Publishing,Paris. 2.2

OECD (2018). Effective Teacher Policies: Insights from PISA. PISA, OECD Publishing, Paris. 1

Sánchez, C. (2018). Targeted or Universal? Mobilizing Students Through School Vouchers. 4

Santiago, P. (2002). Teacher Demand and Supply: Improving Teaching Quality and AddressingTeacher Shortages. Technical Report 1, Paris. 3, 1

Santiago, P., Benavides, F., Danielson, C., Goe, L., and Nusche, D. (2013). OECD Reviews ofEvaluation and Assessment in Education Teacher Evaluation in Chile 2013. OECD Publishing.2.2

Schacter, J. and Thum, Y. M. (2004). Paying for high- and low-quality teaching. Economics ofEducation Review, 23(4):411–430. 1

Seng Tang, O. (2015). Teacher Polices: Global Best Practices for Developing the Teaching Profession.National Institute of Education, Singapore. 2, 4

Solis, A. (2017). Credit Access and College Enrollment. Journal of Political Economy, 125(2):562–622. 4

Sutcher, L., Darling-Hammond, L., and Carver-Thomas, D. (2016). A Coming Crisis in Teaching?Teacher Supply, Demand, and Shortages in the U.S. Technical report, Learning Policy Institute,Palo Alto, CA. 1

Xu, Z., Hannaway, J., and Taylor, C. (2011). Making a difference? The effects of Teach For Americain high school. Journal of policy analysis and management, 30(3):447–469. 2

40

Appendix

A Evolution of Tuition Aid

Figure AF1: Evolution of Tuition Aid: 2000-2016

Source: Mi Futuro - Chile. Amount in 2016 includes funds spent through the free-college policy. Owncalculations.

41

B Income Deciles Used in Scholarship Assignment Pro-

cesses

Table A1: Income Deciles Used in 2015 Scholarship Assignment Process

Income Lower bound Upper bounddecile (in CLP) (in CLP)1 - 48,750

2 48,752 74,969

3 74,970 100,709

4 100,710 125,558

5 125,559 154,166

6 154,167 193,104

7 193,105 250,663

8 250,664 352,743

9 352,744 611,728

10 611,729 upSource: http://portales.inacap.cl

42

C Percent of Recipients of Tuition Scholarships: 2015 -

2016

Table A2: Percent of Recipients of Tuition Scholarships by Mother’s Education: 2015-2016

Programs in 2015 2016S ≤ 12 S = 12 12 < S < 16 S ≥ 16 S ≤ 12 S = 12 12 < S < 16 S ≥ 16

EducationFree tuition 0% 0% 0% 0% 74% 62% 40% 19%BVP 12% 17% 23% 24% 5% 9% 14% 21%BBIC 64% 56% 38% 22% 8% 11% 11% 9%BEA 1% 1% 0% 1% 1% 1% 0% 1%BJGM 10% 10% 9% 5% 3% 5% 6% 4%Other 0% 1% 1% 4% 0% 0% 1% 3%

Non-EducationFree tuition 0% 0% 0% 0% 71% 57% 33% 15%BVP 1% 1% 1% 1% 0% 0% 1% 1%BBIC 63% 56% 35% 19% 8% 10% 10% 7%BEA 3% 3% 2% 2% 2% 2% 2% 1%BJGM 12% 13% 11% 5% 8% 9% 9% 4%Other 0% 1% 1% 5% 0% 1% 1% 4%

Tables A2 and A3 present descriptive evidence further supporting the use of mother’s ed-

ucation and school type as proxies of family income. The percentage of awardees of the

free-tuition grant is higher for students with relatively less-educated mothers, as well as for

those coming from public and voucher schools. Additionally, between 2015 and 2016, the

percentage of BVP awardees decreased more for those with relatively less-educated mothers

and/or coming from public and voucher schools, while the total proportion of scholarship

recipients increased (mostly through the free-tuition grant). This is also the case if we look

at the substitutions patterns between the other two large scholarship programs (i.e., BBIC

and BJGM) and the free-tuition grant, both across education and non-education majors. To

sum up, we see that groups of students whom we consider as relatively poorer are recipients

43

of the free-tuition grant and substitute out of economically dominated grants into free-tuition

status at a larger extent than those who we label as relatively wealthier.

Table A3: Percent of Recipients of Tuition Scholarships by School Types: 2015-2016

Programs in 2015 2016Public Voucher Private Public Voucher Private

EducationFree tuition 0% 0% 0% 66% 54% 10%BVP 17% 18% 26% 9% 10% 23%BBIC 60% 50% 9% 10% 11% 2%BEA 1% 1% 0% 1% 1% 0%BJGM 7% 11% 4% 3% 6% 3%Other 1% 1% 0% 1% 1% 1%

Non-EducationFree tuition 0% 0% 0% 63% 49% 7%BVP 1% 1% 0% 0% 0% 0%BBIC 61% 49% 8% 10% 11% 3%BEA 3% 3% 0% 2% 2% 0%BJGM 11% 13% 4% 7% 10% 4%Other 1% 2% 1% 1% 2% 1%

44

D Eligibility Criteria for Beca Bicentenario (BBIC) and

Beca Juan Gomez Millas (BJGM)

Table A4: Eligibility Criteria for Beca Bicentenario (BBIC) and Beca Juan Gomez Millas(BJGM)

Year Decree Beca Beca JuanBicentenario (BBIC) Gomez Millas (BJGM)

2011Decreto 39 To study at CRUCH universities To study at CRUCH or

de 2011 quality-accredited institutions1-2 income quintiles 1-2 income quintiles

PSU ≥ 550 PSU ≥ 550

2012Decreto 116 To study at CRUCH universities To study at CRUCH or

de 2012 quality-accredited institutions1-3 income quintiles 1-3 income quintiles

PSU ≥ 550 PSU ≥ 550

2013Decreto 97 To study at CRUCH universities To study at CRUCH or

de 2013 quality-accredited institutions1-7 income deciles 1-7 income deciles

PSU ≥ 500 PSU ≥ 500

2014Decreto 167 To study at CRUCH universities To study at CRUCH or

de 2014 quality-accredited institutions1-7 income deciles 1-7 income deciles

PSU ≥ 500 PSU ≥ 500

2015Decreto 108 To study at CRUCH universities To study at CRUCH or

de 2015 quality-accredited institutions1-7 income deciles 1-7 income deciles

PSU ≥ 500 PSU ≥ 500

2016Decreto 253 To study at CRUCH universities To study at CRUCH or

de 2016 quality-accredited institutions1-7 income deciles 1-7 income deciles

PSU ≥ 500 PSU ≥ 500

45

E Tuition-Free College and the Extensive Margin

Table A5: Effect of Free College on Program Vacancies and First-Year Enrollment

(1) (2) (3) (4) (5)Closing Vacancies Freshmen Vacancies FreshmenPrograms Enrollment Enrollment

Post 0.003 0.475 0.530 -0.711 -0.813(0.015) (1.496) (1.750) (4.416) (4.949)

(3.839) (4.159)Post#Agric. 0.839 1.679

(8.381) (9.905)Post#Arts 1.697 2.771

(7.712) (8.998)Post#Scien. 0.965 2.748

(8.462) (9.955)Post#Soc.Scien. 2.547 3.276

(6.674) (7.800)Post#Law 0.802 2.632

(10.462) (12.516)Post#Educ. 0.297 2.483

(5.343) (6.108)Post#Human. 4.748 5.990

(11.813) (13.983)Post#Health 0.698 1.288

(5.722) (6.624)Post#Tech. 1.249 -0.284

(5.131) (5.769)Constant 0.168*** 49.931*** 54.830*** 62.135*** 67.389***

(0.011) (1.080) (1.246) (3.334) (3.587)

Observations 2,591 2,014 2,152 2,014 2,152Note: All regressions area of study fixed-effects. Standard errors in parentheses. Vacanciesand junior rnrollment regressions ran in a sample conditioned on being a non-closing program.*p < 0.1, **p < 0.05, ***p < 0.05.

46

F Checking the Parallel Trends Assumption

Table A6: Difference in Applications to Teaching Majors for Students with PSU ≥ 600Relative to Students with PSU < 600 by School Type (2015 Baseline Year)

Teaching as a Choice Teaching as Top ChoiceCoef. Std.Err. Coef. Std.Err.

Public2014 -0.009 (0.011) -0.009 (0.007)2016 -0.024*** (0.009) -0.010* (0.006)

Voucher2014 -0.004 (0.006) -0.001 (0.004)2016 -0.010* (0.006) -0.004 (0.004)

Private2014 -0.006 (0.007) -0.007 (0.006)2016 -0.002 (0.007) -0.003 (0.005)

Obs. 223,846 223,846Note: We present linear combination of the relevant coefficients for each cate-gory. All regressions include comuna fixed-effects, gender, and linear PSU scorecontrols. The 600 threshold for the PSU score was chosen based on the mini-mum score required to apply for BVP scholarship. Standard errors clustered atthe school level. *p < 0.1, **p < 0.05, ***p < 0.05.

The validity of the difference-in-difference strategy we use in this paper relies on the assump-

tion that the application behavior among students in the treatment and control groups had

the same trend before treatment ensued. We test this crucial assumption using data from

2014 college applications, one year before the timeframe of our analysis. Therefore, we ex-

pand our specification to include coefficients not only for 2016 (i.e., Post = 1 in the original

specification) but also 2014. The regression becomes:

Pi,t =G−1∑g=0

2015∑t=2014

(βgY

gi,tt+ γgY

gi,t1[PSUi,t ≥ 600] + δgY

gi,tt1[PSUi,t ≥ 600]

)+Xi,tθ + µs + εi,t

47

If the pre-treatment trends assumption holds, and given that we treat 2015 as the baseline

year, the coefficients associated with 2014 should not be statistically different from 0.

Table A7: Difference in Applications to Teaching Majors for Students with PSU ≥ 600Relative to Students with PSU < 600 by Mother’s Education (2015 Baseline Year)

Teaching as a Choice Teaching as Top ChoiceCoef. Std.Err. Coef. Std.Err.

S<122014 -0.011 (0.014) 0.001 (0.010)2016 -0.022* (0.013) 0.006 (0.008)

S=122014 -0.004 (0.008) -0.005 (0.006)2016 -0.014** (0.007) -0.010** (0.005)

12<S<162014 0.008 (0.008) 0.000 (0.006)2016 -0.006 (0.008) -0.004 (0.006)

S ≥ 162014 -0.012 (0.008) -0.004 (0.006)2016 -0.011 (0.008) -0.005 (0.005)

Obs. 204,487 204,487Note: We present linear combination of the relevant coefficients for each cate-gory. All regressions include school fixed-effects, gender, and linear PSU scorecontrols. The 600 threshold for the PSU score was chosen based on the mini-mum score required to apply for BVP scholarship. Standard errors clustered atthe school level. *p < 0.1, **p < 0.05, ***p < 0.05.

48

G “Reshuffle” versus “New Entrants”

Figure AF2: PSU Score and New Applicants

(a) Fraction of 2016 Applicants versus 2015 Applicantsby PSU Score

5% 10% 25%

.5.5

25

.55

.57

5.6

.62

5.6

5.6

75

% o

f 2

01

6 A

pp

lica

nts

350 400 450 500 550 600 650 700 750PSU Score

(b) Distributions of Propensity Scores

0.0

01

.00

2.0

03

.00

4.0

05

De

nsity

200 400 600 800PSU

2015 2016 2016 PS Subsample

Note: The left panel shows a linear polynomial regression where the dependent variable is takes the value of 1 if the studentis a 2016 applicant and 0 if she is a 2015 applicant. The running variable is the PSU score. The range plot represents the95% confidence intervals. The vertical lines show the location of the 5th, 10th, and 25th percentiles. The right panel shows thePSU score distributions for different samples, namely, the PSU score distribution for college applicants in 2015, 2016, and thesubsample of 2016 after dropping the observations that fall outside the propensity score’s common support where “treatment”is having applied in 2016 and the predicting variables are gender, PSU score, school type, and municipality.

The introduction of free college in 2016 induced many new applicants, students who thought

that the elimination of tuitions would ease their way into college. To see that, note that

college applications from recent high school graduates increased by 3,201 students between

2014 and 2015, years with no access-to-college policy innovations. That represents a year-

to-year increase of 4.58%. The introduction of free college in 2016 increased applications by

8,819 recent high school graduates year-to-year. That represents a relative increase of 12.%.

That is, applications’ growth after the introduction of free college more than doubled the

existing trend. This inflow of new applicants effectively changed the pool of students we

compare pre and post treatment, which could cast a cloud over the validity of our empirical

strategy.

In this appendix, we show our results are robust to addressing the differences in the pool of

applicants. First, note that Chilean net enrollment in tertiary education is extremely high:

49

90.3%. Thus, the room for expansions in applications is limited. In fact, Figure AF2a shows

that if we pool 2015 and 2016 applications by PSU score, we find that 2016 applications are

over-represented at the lowest quantiles of the PSU score distribution. If extra-applicants

were evenly distributed across the PSU score range, Figure AF2a would be a flat line close

to 50%. But that is not the case. For instance, 6 out of every 10 2015 and 2016 applications

that come from the lower 5% on the PSU score distribution are from 2016 students. Such

over-representation of 2016 applicants goes until the 25th percentile of the PSU score, after

which the line flattens to the expected level that results from natural cohort growth.

The over-representation of 2016 applicants in low quantiles of the PSU score obviously trans-

lates to a shift to the left of the 2016 applicants’ PSU distribution relative to that of 2015

applicants, which is evident in Figure AF2b. To address the possible effect such a change

in the pool of applicants may have on our results, we define a subsample of 2016 in which

students were likely to have applied for college even in the absence of the tuition-free policy.

To do so, we run a version of a propensity score matching where the “treatment” is having

applied to college in 2016 versus 2015. As predicting variables we use gender, PSU score,

school type, and municipality. Thus, we capture the differences in those variables between

the applicants in each year. We then apply a nearest neighbor matching and, thus, we drop

the 2016 applicants that are most dissimilar to those in 2015, at least in those observable

domains. We end with a sample with the exact same size in both years.

Figure AF2b shows that keeping only the 2016 applicants who fall within the 2015 common

support drops applicants from the left tail of the PSU distribution. In fact, the remaining

2016 sample—labelled “2016 PS Subsample” in the Figure—has a PSU distribution that

mimics almost perfectly the one from 2015 applicants. We then run the same empirical

strategy as the one that produce results Tables 5 and 6. In Tables A8 and A9, we show the

estimates remain unchanged. Thus, our original estimates and the subsequent insight we

infer from them do not hinge on the inflow of applicants in 2016 who would not have applied

50

Table A8: Effect of Free College on Application to Teaching Majors by School Type

Teaching as A Choice Teaching as Top ChoiceCoef. Std.Err. Coef. Std.Err.

PublicPSU < 600 0.003 (0.006) -0.006 (0.004)PSU ≥ 600 -0.024*** (0.008) -0.016*** (0.004)Diff -0.026*** -0.010*

VoucherPSU < 600 -0.003 (0.004) -0.007*** (0.003)PSU ≥ 600 -0.014*** (0.005) -0.012*** (0.003)Diff -0.010* -0.004

PrivatePSU < 600 -0.001 (0.007) 0.000 (0.005)PSU ≥ 600 -0.001 (0.003) 0.000 (0.002)Diff 0.000 -0.001

Obs. 145504 145504Avg. dep. variable 0.171 0.086

Note: We present the size of the marginal effect for each category. All regressions includecomuna fixed-effects, gender, and linear PSU score controls. The 600 threshold for the PSUscore was chosen based on the minimum score required to apply for BVP scholarship. Standarderrors clustered at the school level. *p < 0.1, **p < 0.05, ***p < 0.05.

in the absence of the tuition-eliminating policy.

Table A9: Effect of Free College on Application to Teaching Majors by Mother’s Schooling

Note: We present the size of the marginal effect for each category. All regressions include schoolfixed-effects, gender, and linear PSU score controls. The 600 threshold for the PSU score waschosen based on the minimum score required to apply for BVP scholarship. Standard errorsclustered at the school level. *p < 0.1, **p < 0.05, ***p < 0.05.

51

H Free College and Academic Proficiency of Those Who

Get to College

Table A10: Effect of Free College on SIMCE Math Scores of Accepted College Applicants

Panel A: OverallN=110935 Other Education Diff.

Post -1.782*** (0.317) -3.441*** (0.630) -1.659**

Panel B: School TypeN=110719 Other Education Diff.

Post#Public -2.685*** (0.724) -4.952*** (1.119) -2.267*Post#Voucher -2.585*** (0.410) -4.167*** (0.797) -1.582*Post#Private -0.314 (0.654) 1.083 (1.737) 1.397

Panel C: Mother’s EducationN=100814 Other Education Diff.

Post#S < 12 -3.179*** (0.821) -3.752** (1.511) -0.573Post#S=12 -2.682*** (0.513) -4.655*** (1.006) -1.973*Post#12 < S < 16 -1.203** (0.571) -4.586*** (1.327) -3.383**Post#S ≥ 16 0.111 (0.588) 0.173 (1.512) 0.062

Note: All regressions include gender controls. Regressions in Panel A and C include school fixed-effects. Regressions in Panel B include comuna fixed-effects. Column Diff presents the differencebetween the point estimates of Education and Other and tests for whether they are statisticallydifferent from each other or not. Standard errors clustered at the school level in parentheses.*p < 0.1, **p < 0.05, ***p < 0.05.

We estimate the same model as in Table 9 but on scores obtained when the students were

in 4th grade of elementary school, around the age of 10.23 That is, eight years before they

consider their tertiary education decision. Table A10 shows result qualitative very similar

to the ones we report using the PSU in Table 9. The students who eight years later end

up in teaching majors scored 8% of a standard deviation less in their 4th grade math test.

Our results when comparing across socioeconomic traits also hold. As is the case with the

PSU, Table A10 documents that the negative selection that affected the teaching majors was23These yearly national tests, called SIMCE, are part of an information system stablished by the Chilean

government to periodically evaluate learning outcomes across the country. They are mandatory and consis-tently evaluate math and language (Spanish) proficiency.

52

greater among poorer students.24

I Effect of of Free College on PSU Score of Enrollees

Table A11: Effect of Free College on PSU Score of Enrollees

Overall School Type Mother’s EducationPSU PSU PSU

Post Post#Public Post#S < 12Other -0.314 Other -4.850*** Other -3.026***

(0.523) (1.153) (1.120)Education -1.956** Education -5.562*** Education -5.130***

(0.523) (1.459) (1.770)Post#Voucher Post#S=12

Other 0.550 Other -0.917(0.668) (0.817)

Education -1.820* Education -2.747**(1.106) (1.313)

Post#Private Post#12 < S < 16Other 1.440 Other 1.232

(1.048) (0.989)Education 3.373 Education -0.885

(2.672) (1.789)Post#S ≥ 16

Other 0.399(0.959)

Education 1.737(2.219)

Obs 90109 89947 81755Note: All regressions include school fixed-effects and gender controls. Standard errors clustered at theschool level in parentheses. *p < 0.1, **p < 0.05, ***p < 0.05.

24The results hold regardless if we consider math scores like in Table A10 or language scores, omitted forbrevity purposes, but available upon request.

53

J Full Regression Tables

Table A12: Effect of Free College by Type of School

(1) (2)Teaching as a Choice Teaching as Top Choice

Voucher -0.019*** -0.006(0.005) (0.004)

Private -0.076*** -0.040***(0.008) (0.005)

Post 0.002 -0.006(0.006) (0.004)

Voucher X Post -0.006 -0.002(0.007) (0.005)

Private X Post -0.001 0.008(0.009) (0.006)

PSU>600 0.003 0.006(0.007) (0.005)

Voucher X PSU>600 -0.014* -0.003(0.008) (0.006)

Private X PSU>600 -0.008 0.007(0.009) (0.006)

Post X PSU>600 -0.025*** -0.010*(0.009) (0.006)

Voucher X Post X PSU>600 0.015 0.006(0.011) (0.007)

Private X Post X PSU>600 0.023** 0.007(0.011) (0.008)

psu -0.001*** -0.000***(0.000) (0.000)

Male -0.043*** -0.029***(0.002) (0.002)

N 154,277 154,277Standard errors in parentheses* p<0.1, ** p<0.05, *** p<0.01

54

Table A13: Effect of Free College by Mother’s Schooling

(1) (2)Teaching as a Choice Teaching as Top Choice

S = 12 0.004 -0.001(0.005) (0.004)

12 < S < 16 0.001 -0.006(0.006) (0.005)

S ≥ 16 -0.016** -0.020***(0.007) (0.005)

Post -0.005 -0.016***(0.006) (0.004)

S = 12 X Post 0.001 0.010*(0.007) (0.006)

12 < S < 16 X Post 0.002 0.013**(0.008) (0.006)

S ≥ 16 X Post 0.010 0.018***(0.009) (0.007)

PSU>600 0.002 -0.003(0.010) (0.007)

S = 12 X PSU>600 -0.010 0.006(0.011) (0.008)

12 < S < 16 X PSU>600 -0.015 0.006(0.011) (0.008)

S ≥ 16 X PSU>600 -0.003 0.015*(0.011) (0.008)

Post X PSU>600 -0.022* 0.005(0.013) (0.008)

S = 12 X Post X PSU>600 0.007 -0.015(0.015) (0.010)

12 < S < 16 X Post X PSU>600 0.015 -0.009(0.015) (0.010)

S ≥ 16 X Post X PSU>600 0.010 -0.011(0.015) (0.010)

psu -0.001*** -0.000***(0.000) (0.000)

Male -0.037*** -0.026***(0.003) (0.002)

N 140,984 140,984Standard errors in parentheses* p<0.1, ** p<0.05, *** p<0.01

55